fix(runtime): strip module shebang, stream large http responses, reject oversized vm.fetch by content-length

[NathanFlurry]

Summary

Three runtime fixes (plus a limits-inventory classification) needed to unblock Claude agent sessions and large HTTP responses in the VM.

1. Shebang stripping (v8-runtime): build_module_source now strips a leading #! shebang line before compiling a module. The patched Claude agent SDK begins with #!/usr/bin/env node; host Node strips it, but the guest V8 module loader did not, so the adapter SyntaxError-crashed on session/new. Adds a unit test (strip_leading_shebang_matches_node).
2. Streaming large HTTP responses (v8-bridge): ServerResponseBridge now streams a single res.end(body) on a socket-backed server to the connection socket in 256 KiB slices (computing byte length incrementally), instead of buffering the whole body plus its serialize/transmit copies — which tripped the per-isolate heap-limit OOM guard on multi-MB responses. Socket-backed path only; loopback HTTP is unchanged.
3. Oversized vm.fetch rejection (sidecar): parse_kernel_http_fetch_response rejects a response whose declared Content-Length exceeds VM_FETCH_BUFFER_LIMIT_BYTES before stalling on a body the guest cannot stream through the bounded 4 MiB kernel socket buffer before the request deadline. Uses the hard wire limit, so smaller-but-over-a-configured-limit responses still deliver and hit the existing payload-size check.
4. limits-inventory: classify DEFAULT_HEAP_LIMIT_MB (policy, wired to VmLimits.js_runtime.v8_heap_limit_mb) and MAX_V8_USERLAND_CODE_BYTES (invariant).

Testing

• cargo clippy --workspace --all-targets -- -D warnings ✅
• cargo fmt --check ✅
• limits_audit 2/2 ✅
• strip_leading_shebang_matches_node ✅
• All 8 vm_fetch_kernel_tcp_* tests + http-server tests (incl. normal delivery aae) ✅ — aai oversized-rejection now passes, no regressions
• Downstream: agent-os claude-session 9/9 + pi-vanilla-bash 3/3 against this build ✅